kky
20hr ago
In other news: a streamer with deep pockets and a love of AI has decided to have Claude play Pokemon.
To get this working, ClaudeFan (as I'll be calling the anonymous streamer) set up some fairly sophisticated architecture: in addition to the basic I/O shims required to allow an LLM to interface with a GameBoy emulator and a trivial pathfinder tool, Claude gets access to memory in the form of a "knowledge base" which it can update as it desires and (presumably) keep track of what's happening throughout the game. All this gets wrapped up into prompts and sent to Claude 3.7 for analysis and decision. Claude then analyzes this data using a <thinking>reasoning model</thinking>, decides on its next move, and then starts the process over again. Finally, while ClaudeFan claims that "Claude has no special training for Pokemon," it's obvious by the goal-setting that the AI has some external knowledge of where it's supposed to go - it mentions places that it has not yet reached by name and attempts to navigate towards them. Presumably part of Claude's training data came from GameFaqs. (Check out the description on the Twitch page for more detail on the model.)
So, how has this experiment gone?
In a word: poorly. In the first week of playing, it managed to spend about two days wandering in circles around Mt Moon, an early-game area not intended to be especially challenging to navigate. It managed to leave after making a new decision for unexplained reasons. Since then, it has been struggling to navigate Cerulean City, the next town over. One of its greatest challenges has been a house with a yard behind it. It spent some number of hours entering the house, talking to the NPC inside, exhausting all dialogue options, going out the back door into the yard, exploring the yard thoroughly (there are no outlets), re-entering the house, and starting from the top. It is plausible, though obviously not possible to confirm, that ClaudeFan has updated the model some to attempt to handle these failures. It's unclear whether these updates are general bugfixes
How should we interpret this? On the simplest level, Claude is struggling with spacial modeling and memory. It deeply struggles to interpret anything it's seeing as existing in 2D space, and has a very hard time remembering where it has been and what it has tried. The result is that navigation is much, much harder than we would anticipate. Goal-setting, reading and understanding dialogue, and navigating battles have proven trivial, but moving around the world is a major challenge.
The current moment is heady for AI, specifically LLMs, buoyed up by claims by Sam Altman types of imminent AGI. Claude Plays Pokemon should sober us a little to that. Claude is a top performer on things like "math problem-solving" and "graduate-level reasoning", and yet it is performing at what appears to me below the first percentile at completing a video game designed for elementary schoolchildren. This is a sign that what Claude, and similar tools, are doing is not in fact very analogous to what humans do. LLM vendors want the average consumer to believe that their models are reasoning. Perhaps they are not doing that after all?
It's a bit of a tired point, but LLMs are known to be "next likely text" generators. Given textual input, they predict the most likely desired output and return it. Their power at doing this is quite frankly superhuman. They can generate text astonishingly quickly and with unparalleled flexibility in style and capacity for word use. It appears that they are so good at handling this that they are able to pass tests as if they were actually reasoning. The easiest way to trip them up, on the other hand, is to give them a question that is very much like a very common question in their training data but with an obvious difference that makes the default answer inappropriate. The AI will struggle to get past its training and see the question de novo, as a human would be able to. (In case anyone remembers - this is the standard complaint that AI does not have a referent for any of the words it uses. There is no model outside of the language.)
So, as you might guess, I'm pretty firmly on the AI-skeptic side as far as LLMs are concerned. This is usually where these conversations end, as the AI-skeptics believe they've proven their case and (as I understand it) the AI-optimists don't believe that the skeptics have any kind of provable, or even meaningful, model for what intelligence is. But I do actually believe that AGI (meaning: AI that can reason generally, like a human - not godlike Singularity intelligence) is possible, and I want to give an account of what that would entail.
First, and most obviously, an actual AGI must be able to learn. All our existing AI models have totally separate learning and output phases. This is not how any living creature works. An actual intelligence must be able to learn as it attempts to apply its knowledge. This is, I believe, the most natural answer for what memory is. Our LLMs certainly appear to "remember" things that they encountered during their training phase - the fault is in our design that prevents them from ever learning again. However, this creates new problems in how to "sanitize" memory to ensure that you don't learn the wrong things. While the obvious argument around Tay was whether it was racist or dangerously based, a more serious concern is: should an intelligence allow itself to get swayed so easily by obviously biased input? The users trying to "corrupt" Tay were not representative and were not trying to be representative - they were screwing with a chatbot as a joke. Shouldn't an intelligence be able to recognize that kind of bad input and discard it? Goodness knows we all do that from time to time. But I'm not sure we have any model for how to do that with AI yet.
Second, AI needs more than one capacity. LLMs are very cool, but they only do one thing - manipulate language. This is a core behavior for humans, but there are many other things we do - we think spacially and temporally, we model the minds of other people, we have artistic and other sensibilities, we reason... and so on. We've seen early success in integrating separate AI components, like visual recognition technology with LLMs (Claude Play Pokemon uses this! I can't in good faith say "to good effect," but it does open meaningful doors for the AI). This is the direction that AGI must go in.
Last, and most controversial: AI needs abstract "concepts." When humans reason, we often use words - but I think everyone's had the experience of reasoning without words. There are people without internal monologues, and there are the overwhelming numbers of nonverbal animals in the world. All of these think, albeit the animals think much less ably than do humans. Why, on first principles, would it make sense for an LLM to think when it is built on a human capability absent in other animals? Surely the foundation comes first? This is, to my knowledge, completely unexplored outside of philosophy (Plato's Forms, Kant's Concepts, to name a couple), and it's not obvious how we could even begin training an AI in this dimension. But I believe that this is necessary to create AGI.
Anyway, highly recommend the stream. There's powerful memery in the chat, and it is VERY funny to see the AI go in and out of the Pokemon center saying "Hm, I intended to go north, but now I'm in the Pokemon center. Maybe I should leave and try again?" And maybe it can help unveil what LLMs are, and aren't - no matter how much Sam Altman might wish otherwise!
Ha, wow! Was not aware of that. I guess that makes sense w/r/t the funding.
You've written a lot. I think it's best to focus. (As much as I'm tempted to talk about concepts.)
What I understand to be your main point is (my words because you did not state it in concrete terms):
Which is a fair point! The only counterargument to that is on the specifics: why is it improving and what do we expect future improvements to look like? Almost all of the improvement thus far is based on throwing more compute at the problem - so if we're going to see improvements of the same kind, we should see them based on more compute. However, improvements in models are logarithmic - steps up in capacity tend to require 10x compute (by appearances you're pretty educated about AI, so I suspect that is not news to you). So although improvements in efficiency can effectively allow for somewhat more compute, like with Deepseek, we should expect that throwing more compute at the problem will get prohibitively expensive. I believe this has already happened. So while under hypothetical conditions of infinite compute we could have an LLM that infinitely approximates an AGI, similar to the implausible premise of Searle's Chinese Room (a book that allows one to construct a correct response to any input), we are unlikely to see that in practice.
So, how are we to get to AGI, in my opinion? By improving AI on completely different parameters from what currently exists - a revolution in thought about how AI should function. And tests like Claude Plays Pokemon are a fun way of showing us where the gaps in our thinking are.
For my own point of view:
That's not the argument. The argument is: this AI is struggling in a VERY non-human way with what we would consider a pretty trivial task. This reveals that its operational parameters are not like those of a human, and that we should figure out where else it is going to perform at sub-human levels. The fact that we're seeing this at the same time as it performs at SUPERhuman levels in other tasks shows that this is not AGI, or even in the direction of AGI, but rather is tool AI. (I assume you think humans are, at the very least, general intelligence - right?)
I don't think you've addressed this point, except here:
Why should I care? AGI is supposed to be GENERAL. This is the stuff that's supposed to be taking people's jobs in a few years! And yet it gets lost in Cerulean City? As a tech demo, this is very cool - it's remarkable that someone was able to pipe these pieces together, and the knowledge base idea is very cool and is a plausible direction to take new LLMs into. A hypothetical Claude 3.8 that is explicitly trained to make knowledge base manipulation a central feature of the model could potentially perform miles better on some of these tasks. But all you've told me is that I should expect AI to struggle with these tasks. In which case: doesn't it sound like we agree? We both agree that there was no reason to expect Claude to succeed with Pokemon at the level of an eight-year-old. So, from the perspective of an uncommitted third party, given that an AI skeptic and an AI optimist have both agreed that an LLM can't play Pokemon like an eight-year-old... well, it feels pretty clear to me.
Obviously, if this becomes a big selling point for the next generation of LLMs, then we'll see them all benchmarked on Pokemon Red speedruns and you can I-told-you-so about AI being able to beat Pokemon. I don't doubt the ability of motivated corporations to "teach to the test" - it's what we've been seeing with "reasoning" AIs. It's just one of the problems with setting up real tests of ability for some of these AIs, because they get so much data that it's all but impossible to ensure you have a pure test like what the IQ test aspires to.
More options
Context Copy link